Retinal Disease Classification with Domain Generalization

An end-to-end deep learning research project for retinal disease classification under cross-dataset domain shift.

This project investigates how retinal disease classification models behave when trained on retinal fundus datasets from different sources and evaluated on an unseen dataset/domain.

Rather than relying on a conventional random train-test split from a single dataset, the project explores multiple training strategies, model configurations, and domain generalization approaches to study whether learned representations can transfer across different retinal imaging domains.

The project also includes an interactive Streamlit application for exploring the datasets, experimental approaches, model behavior, and predictions.

---

Overview

Deep learning models for medical image classification often perform well when training and testing data come from the same distribution.

However, real-world medical imaging data can vary considerably due to:

• Different imaging devices
• Different acquisition protocols
• Differences in patient populations
• Different geographic or institutional sources
• Image quality variations
• Dataset-specific characteristics
• Distribution and domain shifts

A model trained on one retinal dataset may therefore struggle when exposed to images from another dataset.

This project focuses on this problem through Domain Generalization (DG).

The main research question is:

How well can retinal disease classification models generalize to an unseen retinal dataset when the training data come from multiple different domains?

---

Research Objectives

The project was designed to investigate:

• Cross-dataset generalization in retinal disease classification
• The effect of training with multiple retinal datasets
• The performance gap between conventional and cross-domain evaluation
• Whether domain-generalization strategies improve robustness to unseen domains
• The effect of different model approaches under domain shift
• The usefulness of hybrid approaches for improving cross-domain representation learning

---

Research Setup

The project uses multiple retinal fundus image datasets originating from different domains.

Instead of treating all images as coming from a single homogeneous distribution, each dataset is considered a separate domain.

General Setup

Multiple Retinal Datasets
          │
          ▼
   Data Harmonization
          │
          ▼
   Preprocessing
          │
          ▼
 ┌─────────────────────┐
 │ Different Learning  │
 │      Approaches     │
 └─────────────────────┘
          │
          ├── Baseline Approach
          │
          ├── Cross-Dataset Evaluation
          │
          ├── Domain Generalization
          │
          └── Hybrid Approach
          │
          ▼
   Unseen-Domain Testing
          │
          ▼
   Cross-Domain Analysis